What Places an Action Boundary? Structure-Aware Attribution for Temporal Action Localization
Organizations: Department of Computer Science, University of Arizona · Silicon Austria Labs · Department of Electrical and Computer Engineering, University of Arizona
Abstract
Temporal action localization (TAL) detectors must identify actions and estimate their temporal boundaries, but localization remains a major source of detection errors. While understanding what determines boundary placement requires faithful attribution of temporal evidence, existing attribution methods have not sufficiently addressed the temporal structure of TAL or the need to tailor attribution to boundary localization. We introduce a structure-aware attribution framework that averages contributions across random evidence-restoration orders following the Shapley-Shubik principle. We make it tractable by using the detector's receptive field to limit which time steps are ordered individually, and by approximating each order with a soft gradient path. For boundary predictions, we further introduce cell-structured restoration orders and boundary-anchored references that selectively remove adjacent evidence. We also adapt faithfulness evaluation to assess the joint effects of contiguous temporal evidence and the direction of boundary displacement. Experiments on THUMOS14 and EPIC-KITCHENS-100 with TriDet and ActionFormer demonstrate improved attribution faithfulness over existing methods. Attribution analysis and controlled interventions further reveal that class-score changes do not reliably track boundary displacement, providing input-level evidence that classification relies on content less informative about boundary placement. Code and evaluation artifacts will be publicly released.
Figures & tables
| TriDet | ActionFormer | |||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| THUMOS14 | EPIC-KITCHENS | Hours | THUMOS14 | EPIC-KITCHENS | Hours | |||||||||||||||||
| Sens.- | Sens.- | Sens.- | Sens.- | |||||||||||||||||||
| Method | Win. | Path | Del. | Ins. | rem. | rest. | Del. | Ins. | rem. | rest. | TH | EP | Del. | Ins. | rem. | rest. | Del. | Ins. | rem. | rest. | TH | EP |
| Single path | ||||||||||||||||||||||
| IG [ 46 ] | 61.2 | 39.8 | 62.4 | 78.0 | 50.3 | 34.3 | 62.1 | 64.4 | – | – | 63.1 | 38.9 | 66.7 | 81.8 | 47.6 | 35.2 | 59.3 | 62.9 | – | – | ||
| Guided IG [ 21 ] | 61.3 | 39.8 | 62.4 | 78.1 | 50.4 | 34.5 | 62.2 | 64.5 | – | – | 63.2 | 38.9 | 66.7 | 81.8 | 47.6 | 35.2 | 59.5 | 63.1 | – | – | ||
| TriDet | ActionFormer | |||||||||||||||||||
| THUMOS14 | EPIC-KITCHENS | Hours | THUMOS14 | EPIC-KITCHENS | Hours | |||||||||||||||
| Ins.– | Sens.- | Ins.– | Sens.- | Ins.– | Sens.- | Ins.– | Sens.- | |||||||||||||
| Method | Win. | Cells | Refs. | Path | Del. | rem. | rest. | Del. | rem. | rest. | TH | EP | Del. | rem. | rest. | Del. | rem. | rest. | TH | EP |
| Single path | ||||||||||||||||||||
| IG [ 46 ] | 22.2 | 35.6 | 43.2 | 15.0 | 13.6 | 25.6 | – | – | 22.2 | 43.4 | 44.1 | 5.5 | 2.4 | 7.6 | – | – | ||||
| Guided IG [ 21 ] | 22.4 | 36.1 | 43.7 | 15.2 | 14.0 | 25.9 | – | – | 22.4 | 43.9 | 44.7 | 5.5 | 2.3 | 7.7 | – | – | ||||
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | Implementation settings |
|---|---|
| IG [ 46 ] | Dataset-mean reference; straight-line path; 100 integration intervals. |
| Guided IG [ 21 ] | Dataset-mean reference; 25% lowest-gradient coordinates per step; maximum deviation 0.02 from the straight line; 100 steps. |
| Adaptive nodes [ 35 ] | Adaptive integration-node placement on the input straight-line path; 100 nodes. |
| Left to right | Restore the context unit, then the window steps one at a time from left to right. |
| Right to left | Restore the context unit, then the window steps one at a time from right to left. |
| TIMING [ 18 ] | 50 masked segments of 10–48 input steps; all feature channels masked together. |
| Class | Boundary | |||||||
| Sens.- | Ins.– | Sens.- | ||||||
| Map | Del. | Ins. | rem. | rest. | Del. | rem. | rest. | |
| TriDet | ||||||||
| TH (176) | TIMING | 60.5 | 40.7 | 65.7 | 75.3 | 23.8 | 46.6 | 42.5 |
| Ours | 63.5 | 44.1 | 68.5 | 81.5 | 25.9 | 44.3 | 48.3 | |
| EP (132) | TIMING | 52.1 | 37.3 | 64.2 | 65.3 | 19.0 | 34.6 | 31.0 |
| (a) Class-score drop, half of replaced | |||||
|---|---|---|---|---|---|
| Contig. | Random | Even | |||
| THUMOS14 | 65.3 | 55.2 | 49.5 | ||
| EPIC-KITCHENS | 54.5 | 46.5 | 38.1 | ||
| (b) Joint block effect / sum of step effects | |||||
| Block width | 2 | 4 | 8 | ||
| THUMOS14 | 1.56 | 1.69 | 1.76 | ||
| THUMOS14 | EPIC-KITCHENS | |||||||
| Sens.- | Sens.- | |||||||
| Orders | Del. | Ins. | rem. | rest. | Del. | Ins. | rem. | rest. |
| TriDet | ||||||||
| Unstructured | 63.0 | 43.2 | 68.3 | 81.4 | 56.3 | 41.3 | 71.1 | 73.7 |
| + window | 63.3 | 43.8 | 68.2 | 81.2 | 56.1 | 41.3 | 70.9 | 73.5 |
| + cells | 62.8 | 43.2 | 67.6 | 81.5 | 55.1 | 38.8 | 69.2 | 71.9 |
| TriDet | ActionFormer | |||||||
| Cells window | Cells window | |||||||
| Level | Exact soft | soft | exact | Exact soft | soft | exact | ||
| 0 | 56 | 0.3 | 0.3 | 0.5 | 57 | 0.2 | 0.9 | 0.6 |
| 1 | 57 | 0.1 | 0.5 | 0.5 | 57 | 0.0 | 1.6 | 1.5 |
| 2 | 54 | 1.2 | 1.0 | 2.4 | 54 | 2.9 | 2.9 | 4.3 |
| 3 | 53 | 3.5 | 6.0 | 8.5 | 53 | 5.5 | 7.5 | 8.5 |
| Block width (input steps) | |||||||
| Map | 1 | 2 | 4 | 8 | 16 | 32 | |
| THUMOS14, removal | |||||||
| IG | 33.9 | 33.2 | 33.0 | 35.0 | 33.5 | 29.5 | 43.2 |
| Unstructured | 34.6 | 36.0 | 36.4 | 37.8 | 36.0 | 32.1 | 47.5 |
| + cells (ours) | 34.9 | 36.8 | 38.9 | 41.4 | 40.2 | 37.8 | 51.4 |
| + references (ours) | 38.4 | 41.0 | 43.5 | 45.6 | 43.6 | 41.2 | 53.4 |
| vs. cell width | ||||||
|---|---|---|---|---|---|---|
| Dataset | Sensitivity- | One cell | Other | |||
| TriDet | ||||||
| THUMOS14 | removal | 0.9 | 3.9 | 4.8 | 4.5 | 4.5 |
| restoration | 3.2 | 4.1 | 3.9 | 3.5 | ||
| EPIC-KITCHENS | removal | 3.6 | 7.5 | 4.9 | 18.3 | 6.4 |
| restoration | 4.0 | 4.5 | 9.3 | 4.1 | ||
| THUMOS14 | EPIC-KITCHENS | |||||
| Sens.- | Sens.- | |||||
| Reference | Ins.–Del. | rem. | rest. | Ins.–Del. | rem. | rest. |
| Straight line | ||||||
| Dataset mean (IG) | 22.2 | 35.6 | 43.2 | 15.0 | 13.6 | 25.6 |
| Random spans in | 22.7 | 39.5 | 41.6 | 19.0 | 25.2 | 28.8 |
| Boundary spans | 23.9 | 41.7 | 43.5 | 19.4 | 27.4 | 32.0 |
| TH | EP | IV | AF | |
| Strongest expanding step in the boundary’s cell | ||||
| or the next cell inside (%) | 59 | 80 | 73 | 68 |
| Class positive at the boundary’s cell (%) | 87 | 85 | 78 | 87 |
| and one cell beyond it (%) | 71 | 66 | 49 | 67 |
| The steps just inside the boundary | ||||
| Share of the near half’s steps (%) | 33 | 35 | 33 | 33 |
| Boundary map | TH | EP | IV | AF | |
|---|---|---|---|---|---|
| Inside | Without references | 0.16 [ 0.26, 0.06] | 0.11 [ 0.20, 0.01] | 0.14 [ 0.24, 0.02] | 0.01 [ 0.10, 0.11] |
| partial, | 0.11 [ 0.21, 0.01] | 0.15 [ 0.25, 0.05] | 0.10 [ 0.20, 0.01] | 0.06 [ 0.04, 0.15] | |
| partial, and level | 0.11 [ 0.21, 0.01] | 0.17 [ 0.27, 0.07] | 0.09 [ 0.19, 0.01] | 0.06 [ 0.03, 0.16] | |
| Fused | 0.07 [ 0.18, 0.04] | 0.17 [ 0.25, 0.07] | 0.06 [ 0.06, 0.19] | 0.01 [ 0.11, 0.12] | |
| Outside | Without references | 0.07 [ 0.17, 0.04] | 0.02 [ 0.12, 0.07] | 0.18 [0.07, 0.30] | 0.12 [0.02, 0.22] |
| partial, | 0.09 [ 0.20, 0.02] | 0.06 [ 0.16, 0.03] | 0.18 [0.07, 0.29] | 0.09 [ 0.02, 0.19] |
| Inside the box | Outside the box | |||||||
| TH | EP | IV | AF | TH | EP | IV | AF | |
| Boundary: outward shift per replaced step | ||||||||
| Measured | .46 | .22 | .61 | .46 | .18 | .05 | .20 | .16 |
| Expected | .13 | .05 | .45 | .28 | .07 | +.13 | .24 | .04 |
| Class score: relative change (%) | ||||||||
| Measured | 8.6 | 10.2 | 8.8 | 7.9 | 3.1 | 5.6 | 1.9 | 2.5 |
| Class-score change | TH | EP | IV | AF | |
|---|---|---|---|---|---|
| Inside | 0.46 [ 0.52, 0.36] (104) | 0.33 [ 0.43, 0.22] (117) | 0.67 [ 0.76, 0.51] (79) | 0.44 [ 0.56, 0.39] (94) | |
| to | 0.48 [ 0.56, 0.34] (50) | 0.31 [ 0.47, 0.15] (53) | 0.55 [ 0.68, 0.37] (32) | 0.46 [ 0.56, 0.41] (48) | |
| to | 0.48 [ 0.59, 0.37] (60) | 0.32 [ 0.40, 0.21] (80) | 0.54 [ 0.75, 0.46] (55) | 0.43 [ 0.54, 0.34] (86) | |
| 0.40 [ 0.52, 0.30] (184) | 0.09 [ 0.20, 0.03] (254) | 0.67 [ 0.80, 0.55] (152) | 0.49 [ 0.57, 0.41] (168) | ||
| Outside | 0.15 [ 0.19, 0.10] (161) | 0.04 [ 0.09, 0.01] (158) | 0.18 [ 0.22, 0.14] (162) | 0.14 [ 0.20, 0.09] (185) | |
| to | 0.15 [ 0.19, 0.12] (90) | 0.00 [ 0.10, 0.05] (77) | 0.19 [ 0.25, 0.09] (54) | 0.16 [ 0.24, 0.09] (73) |
Feature Attribution
Feature attribution estimates which input features or internal components contribute to a model’s predictions.
See all 258 papers